An important issue in maintaining an image interpretation system involves the controlling and monitoring of system performance. The system is a complex system comprising different processing components (e.g., image analysis, feature extraction and high-level image interpretation). The quality of the results of one component strongly depends on the quality of a preceding component. Several possible strategies exist for improving system performance.
Control without Feedback (Local Optimization)
The simplest approach is to adjust the performance of each component without considering the others. Each component - segmentation, feature extraction and selection, and interpretation - acts alone. No interaction between them is allowed.
Image segmentation performance may be determined by subjective evaluation of the segmentation result as done by an expert, by calculating the similarity between the original and segmented images, by interclass distances for feature extraction, or by classification error. This strategy has the advantage that the control of the system is simple. Nevertheless, it cannot optimize system performance because only local optimums can be achieved for each single component.
Control with Feedback (Global Optimization)
If after local optimization the performance of a component could not be improved or is not satisfactory, the control algorithm will lead the learning process to the preceding processing component in an attempt to further improve its performance.
This process stops if the first processing component is reached and if no improvement could be established after local optimization.
The logical scheme in Table 1 shows us how control is guided. If the performance of all components is good, no action has to be taken. If the interpretation component’s performance is poor, then its performance needs to be optimized. We assume that it is impossible for a preceding component to perform poorly while its successor components perform well.
Table 1. Logical Scheme of Performance Control Segmentation (S) Feature Extraction (FE) Interpretation (I) Action
Good Good Good No Action
Good Good Poor Optimize I
Good Poor Good Impossible
Good Poor Poor Optimize FE and
examine effects on I
Poor Good Good Impossible
Poor Good Poor Impossible
Poor Poor Good Impossible
Poor Poor Poor Optimize S, then
re-examine the performance of the other components
8 Conclusion
We surveyed special topics associated with a case-based image interpretation system.
From our point of view case-based image interpretation differs in many aspects from other CBR applications that require further investigation. First, more systematic work on special image similarity measures is needed that investigates the measures under different influences that may occur in an image. Next, case representations are required for all the different abstraction levels of an image. Finally, the maintenance and learning strategies must be defined so that they can help to improve the system performance and discover the range of target problems that the system can solve.
We have recently deployed two CBR image interpretation systems. One is installed at the university hospital in Halle; it is used for image segmentation to determine the brain/liquor ratio of the head in a CT image. The second system is used to interpret ultra-sonic SAFT images. In both applications the CBR strategies we used achieved good system performance that satisfied the users and outperformed other systems.
The learning and maintenance facilities installed to date have been particularly well- received.
In summary, we believe that investigations of case-based image interpretation systems can reveal special challenges to both the CBR and computer vision communities, and encourage more people to work on this topic.
Acknowledgements. We would like to thank M. Richter for helpful discussions on this topic. Special thanks to David Aha for his kind advice on this topic and for his helpful comments that significantly improved earlier versions of this paper. The implementation of the case-based image segmentation on a parallel computer has been funded by the EU project (HP-ISIS). The CBR system for Hep-2 cells is funded within in the project LernBildZell by the German Ministry of Economy .
References
1. Jarmulak, J. (1998). Case-based classification of ultrasonic B-Scans: Case-base organisation and case retrieval. In B. Smyth & P. Cunningham (Eds.) Advances in Case- Based Reasoning (pp. 100-111). Berlin: Springer Verlag.
2. Grimnes, M. & Aamodt, A.(1996). A two layer case-based reasoning architecture for medical image understanding, In I. Smith & B. Faltings (Eds.) Advances in Case-Based Reasoning (pp. 164-178). Berlin: Springer Verlag.
3. Perner, P. (1999). An architecture for a CBR image segmentation system. Journal of Engineering Application in Artificial Intelligence, 12(6), 749-759.
4. Micarelli, A. Neri, A., & Sansonetti, G. (2000). A case-based approach to image recognition, In E. Blanzieri & L. Portinale (Eds.) Advances in Case-Based Reasoning (pp.
443-454). Berlin: Springer Verlag.
5. Venkataraman, S., Krishnan, R., & Rao, K.K. (1993). A rule-rule-case based system for image analysis. In M.M. Richter, S. Wess, K.D. Althoff, & F. Maurer (Eds.) First European Workshop on Case-Based Reasoning (Technical Report SFB 314) (pp. 410- 415). Kaiserslautern, Germany: University of Kaiserslautern.
6. Ficet-Cauchard, V., Porquet, C., & Revenu, M. (1999). CBR for the reuse of image processing knowledge: A recursive retrieval/adaption strategy. In K.-D. Althoff, R.
Bergmann, & L.K. Branting (Eds.) Case-Based Reasoning Research and Development (pp. 438-453). Berlin: Springer.
7. Rice, S.V., Nagy, G., & Nartker, T.H. (1999). Optical character recognition: An illustrated guide to the frontier. London: Kluwer.
8. Klette, R., Stiehl, H.S., Viergever, M.A., & Vincken, K.L. (2000). Performance characterization in computer vision. London: Kluwer
9. Perner, P. (1998). Using CBR learning for the low-level and high-level unit of a image interpretation system. In S. Singh (Ed.) Advances in Pattern Recognition (pp. 45-54).
Berlin: Springer Verlag.
10. Perner, P. (1999). An architecture for a CBR image segmentation system. In K.-D.
Althoff, R. Bergmann, & L. K. Branting (Eds.) Case-Based Reasoning Research and Development (pp. 525-535). Berlin: Springer Verlag.
11. Zamperoni, P. & Starovoitov, V. (1995). How dissimilar are two gray-scale images? In Proceedings of the Seventeenth DAGM Symposium (pp.448-455). Berlin: Springer Verlag.
12. Santini, S. & Jain, R. (1999). Similarity measures IEEE Transactions on Pattern Analysis and Machine Intelligence,. 21(9), 871-883.
13. Horikowa, Y. (1996). Pattern recognition with invariance to similarity transformations based on third-order correlations. In Proceedings of Internatinal Confernce on Pattern Recognition´96, IEEE Computer Society Press (pp 200-204).
14. Leitao, F. (1999). A study of string dissimilarity measures in structural clustering. In: S.
Singh (Ed.) Advances in Pattern Recognition (pp. 385-394). Berlin: Springer Verlag.
15. Mehrotra, G. (1993). Similar shape retrieval using a structural feature index. Information Systems, 18 (5), 525-537.
16. Cortelazzo, C., Deretta, G., Mian, G.A., & Zamperoni, P. (1996). Normalized weighted Levensthein distance and triangle inequality in the context of similarity discrimination of bilevel images. Pattern Recognition Letters, 17(5), 431-437.
17. Crouzil, A., Massipo-Pail, L., & Castan, S. (1996). A new correlation criterion based on gradient fields similarity. In Proceedings of Internatinal Confernce on Pattern Recognition´96, IEEE Computer Society Press, (pp. 632-636).
18. Moghadda, Nastar, & Pentland (1996). A Bayesian similarity measure for direct image matching. In Proceedings of Internatinal Confernce on Pattern Recognition´96, IEEE Computer Society Press, (pp. 350-358).
19. Moghadda, Jebra, & Pentland (1998). Efficient MAP/ML similarity matching for visual recognition, In Proceedings of Internatinal Confernce on Pattern Recognition´98, IEEE Computer Society Press, (pp. 876-881).
20. Wilson, D.L., Baddely, A.J., & Owens R.A. (1997). A new metric for gray-scale image comparison, International Journal of Computer Vision, 24(1), 5-19.
21. Messmer, B., & Bunke, H. (2000). Efficient subgraph isomorphism detection: A decomposition approach. IEEE Trans. on Knowledge and Data Engineering, 12(2), 307- 323.
22. van der Heiden, A., & Vossepoel, A. A landmark-based approach of shape dissimilarity.
In Proceedings of the International Conference on Pattern Recognition’99, IEEE Computer Society Press, (pp. 120-124).
23. Perner, P. (1998) Content-based image indexing and retrieval in a image database for technical domains, In H.H.S. Ip & A. Smuelder (Eds.) Multimedia Information Analysis and Retrieval (pp. 207-224). Berlin: Springer Verlag.
24. Voss, A. (Ed.) Similarity Concepts and Retrieval Methods. Fabel Report No. 13, GMD- 1993, ISSN 0942-413.
25. Perner, P. (1998). Image analysis and classification of HEp-2 cells in fluorescent images, Fourteenth International Conference on Pattern Recognition (pp. 1677-1679). Brisbane Australia: IEEE Computer Society Press.
26. Perner, P., Zscherpel, U., & Jacobsen, C. (2000). Comparison between neural networks and decision trees application of machine learning in industrial radiographic testing.
Pattern Recognition Letters, 22/1, 47-54.
27. Perner, P. (1998). Different learning strategies in a case-based reasoning system for image interpretation. In B.Smyth & P. Cunningham (Eds.) Advances in Case-Based Reasoning (pp. 251-261). Berlin: Springer Verlag.
28. Leake, D.B., & Wilson, D.C. (2000). Remembering why to remember: performance- guided case-base maintenance, In E. Blanzieri & L. Portinale (Eds.) Advances in Case- Based Reasoning (pp. 161-172). Berlin: Springer Verlag.
29. Portinale, L., Torasso, P., & Tavano, P. (1999). Speed-up, quality, and competence in multi-modal reasoning In K.-D. Althoff, R. Bergmann, & L.K. Branting (Eds.) Case- Based Reasoning Research and Development (pp. 303-317). Berlin: Springer.
30. Perner, P. (1994), How to use Repertory Grid for Knowledge Acquisition in Image Interpretation. HTWK Leipzig Report Nr. 2.
31. Schmitt, S., Jerusalem, D., & Nett, T. (2000) Representation and execution of questioning strategies for the acquisition of customer requirements, In M.H. Gửker (Ed.) Eighth German Workshop on Case-Based Reasoning, p.23-37, DaimlerChrysler Research and Technology FT3/KL.
32. Smyth, B., & McKenna, E. (1998). Modeling the competence of case-bases. In B. Smyth
& P. Cunningham (Eds.) Advances in Case-Based Reasoning (pp. 208-220). Berlin:
Springer Verlag.
33. Leake, D.B., & Wilson, D.C. (2000). Remembering why to remember: performance- guided case-base maintenance. In E. Blanzieri & L. Portinale (Eds.) Advances in Case- Based Reasoning (pp. 161-172). Berlin: Springer Verlag.
34. Cordella, Goggia, Sansone, Tortorella, & Vento (1999). Graph matching: A fast algorithm and its evaluation. In Proceedings of the International Conference on Pattern Recognition’99, IEEE Computer Society Press (pp. 1582-1584).
35. Cheetham, W. & Graf, J. (1997), Case-Based Reasoning in Color Matching, In: Leake, D.B. & Plaza, E. (Eds.) Case-Based Reasoning Research and Development, (pp. 1- 12).Berlin, Springer Verlag.
36. Bergmann, R. & Wilke, W.(1996), On the Role of Abstraction in Case Based Reasoning, In: Smith, I. & Faltings, B. (Eds.) Advances in Case-Based Reasoning, (pp. 28-43). Berlin, Springer Verlag.
Eva Armengol and Enric Plaza IIIA - Artificial Intelligence Research Institute, CSIC - Spanish Council for Scientific Research, Campus UAB, 08193 Bellaterra, Catalonia (Spain).
{eva, enric}@iiia.csic.es,
Abstract. Reasoning and learning from cases are based on the concept of similarity often estimated by a distance. This paper presentsLAUD, a distance measure that can be used to estimate similarity among rela- tional cases. This measure is adequate for domains where cases are best represented by relations among entities. An experimental evaluation of the accuracy of LAUD is presented for the task of classifying marine sponges.
1 Introduction
Reasoning and learning from cases is based on the concept of similarity. Often similarity is estimated by a distance (a metric) or a pseudo-metric. This ap- proach proceeds by a pairwise similarity comparison of a problem with every precedent case available in a case base; then one case (ork cases) with greatest (greater) similarity is (are) selected. This process is called the retrieval phase in Case-based Reasoning (CBR), and also plays a pivotal role in lazy learning techniques like Instance-based Learning (IBL) and k-nearest neighbor. In clas- sification tasks, the solution class of the problem is inferred from the solution class of the precedent case(s) selected.
However, distance-based approaches to case retrieval are mainly used for propositional cases, i.e. cases represented as attribute-value vectors. We are in- terested in this paper in learning tasks where cases are best represented in a scheme that uses relations among entities. We will call this setting relational case-based learning. One option to achieve case-based learning in a relational setting is to adapt the process of pairwise similarity comparison by defining a distance that works upon relational instances. An example of similarity to be applied in relational cases is that used by RIBL ([7]) where the cases are repre- sented as collections of Horn clauses (see related work on section 5).
We are interested in using cases represented in a relational way usingfeature terms[1]. Feature terms are a generalization of first order terms. In this represen- tation entities are typed by sorts and relations among entities are represented byfeatures. In this paper we introduceLAUD, a new distance measure that we use to estimate the similarity of relational cases represented asfeature terms.
The structure of this paper is the following. In section 2 we introduce the feature term representation. In section 3 we introduce a new similarity for esti-
D.W. Aha and I. Watson (Eds.): ICCBR 2001, LNAI 2080, pp. 44–58, 2001.
c Springer-Verlag Berlin Heidelberg 2001
mating the similarity of cases represented as feature terms. In section 4 we pro- vide some results of the application of the similarity to identify marine sponges.
Finally, we report some related work and the conclusions and future work.
2 Representation of the Cases
Feature Terms(also called feature structures orψ-terms) are a generalization of first order terms. The difference between feature terms and first order terms is the following: a first order term, e.g.f(x, y, g(x, y)) can be formally described as a tree and a fixed tree-traversal order. In other words, parameters are identified by position. The intuition behind a feature term is that it can be described as a labelled graph i.e. parameters are identified by name. A formal definition of feature terms is the following:
Given a signature Σ = hS,F,i (where S is a set of sort symbols that includes ⊥; F is a set of feature symbols; andis a decidable partial order on S such that ⊥is the least element) and a set ϑof variables, we define feature terms as an expression of the form:
ψ::=X:s[f1=. Ψ1. . . fn =. Ψn] (1) whereX is a variable inϑcalled theroot of the feature term,sis a sort inS, the functionroot(X) returns the sort of the root,f1. . . fn are features inF,n≥0, and eachΨi is a set of feature terms and variables. Whenn= 0 we are defining a variable without features. The set of variables occurring inψis noted asϑψ.
Sorts have an informational order relation () among them, where ψψ0 means that ψ has less information than ψ0 ˜Nor equivalently that ψ is more general than ψ0. The minimal element (⊥) is called any and it represents the minimum information. When a feature has unknown value it is represented as having the value any. All other sorts are more specific than any. We restrict ourselves to use sort hierarchies with single inheritance (every sort has only one most specific supersort).
Using therelation, we can introduce the notion ofleast upper bound (lub) commonly used in ILP [9]. The lub of two sorts is the most specific sort gen- eralizing both. As we will explain in section 2.1 the notion of lub will be used to define the anti-unification of two feature terms. Apath π(X, fi) is defined as a sequence of features going from the variable X to the feature fi. The depth of a feature f in a feature term ψ with root X is the number of features that compose the path from the rootX to f, includingf, with no repeated nodes.
Given a particular maximum feature depthk, aleaf feature of a feature term is a featurefisuch that either 1) the depth offiiskor 2) the value offi is a term without features. We callleaves(ψ , k) the set of leaf features of a termψ.
Let us illustrate the concepts introduced above with an example. The feature term of Figure 1 represents the description of a marine sponge. Theroot of this feature term iss364, the sorts are written initalic(for instance,sponge, external- features, growing, . . . ), some features are external-features, ecological-features, megascleres, separable, aspect, etc. Notice that the features ornamentation and
Fig. 1.Representation of a sponge using feature terms.
vertical-tractsare set-valued. The feature leaves ofs364 are the following{body- size, touch, grow, form, peduncle, hollow, osc, fixation, substrate, location, chemical, architecture, smooth-form, acanthose, ornamentation, microscleres, vertical-tracts, transversal-tracts, distribution, separable, aspect, cortex}. An example of path is π(s364,acanthose) that represents the path from the root to the leaf featureacan- those, i.e. the sequence of features (spiculate-skeleton, megascleres, acanthose).
As we have explained above, there is an order relation between sorts. Figure 2 shows the sort/subsort hierarchy for the values of the featuremegascleres. The most general sort allowed for the values of the featuremegascleresismegas-form and there are several subsorts (e.g. triaena, style, calthrop,etc). In turn, some of these subsorts (e.g.triaena, style, tylote) have subsorts. Let us suppose that
Megas-form
Oxea
Calthrop
Tornote
Tylote Subtylote Cladotylote Triaena
Protriaena Anatriaena Plagiotriaena Mesotriaena Orthotriaena Dichotriaena Amphitriaena
Style
Subtylostyle Tylostyle
Level 4 3 2 1 Desma
Lophoalthropa
Rhabdostyle
Spherotylostyle
Depth 1 2 3 4
Fig. 2.Part of the sort hierarchy corresponding to the values of the featuremegascleres.
It is a hierarchy of depth 4. The levels of the hierarchy will be used to determine the sort similarity.
v1=protriaena,v2=anatriaenaandv3=tylostyle. The least upper bound of v1 andv2,lub(v1, v2), istriaena that is the most specific supersort ofprotriaena and anatriaena whereas the lub(v1, v3) is megas-form since protriaena and tylostyle only share the fact that both are kinds of megascleres.
There are two important concepts concerning the representation using fea- ture terms that will be used later for assessing similarity. One of them is the subsumption relation and the other one is the anti-unification operation. Both concepts are explained in the next section.